Popularity of metasurfaces has been continuously growing due to their attractive properties including the ability to effectively manipulate electromagnetic (EM) waves. Metasurfaces comprise optimized geometries of unit cells arranged as a periodic lattice to obtain a desired EM response. One of their emerging application areas is the stealth technology, in particular, realization of radar cross section (RCS) reduction. Despite potential benefits, a practical obstacle hindering widespread metasurface utilization is the lack of systematic design procedures. Conventional approaches are largely intuition-inspired and demand heavy designer’s interaction while exploring the parameter space and pursuing optimum unit cell geometries. Not surprisingly, these are unable to identify truly optimum solutions. In this article, we introduce a novel machine-learning-based framework for automated and computationally efficient design of metasurfaces realizing broadband RCS reduction. Our methodology is a three-stage procedure that involves global surrogate-assisted optimization of the unit cells, followed by their local refinement. The last stage is direct EM-driven maximization of the RCS reduction bandwidth, facilitated by appropriate formulation of the objective function involving regularization terms. The appealing feature of the proposed framework is that it optimizes the RCS reduction bandwidth directly at the level of the entire metasurface as opposed to merely optimizing unit cell geometries. Computational feasibility of the optimization process, especially its last stage, is ensured by high-quality initial designs rendered during the first two stages. To corroborate the utility of our procedure, it has been applied to several metasurface designs reported in the literature, leading to the RCS reduction bandwidth improvement by 15%–25% when compared with the original designs. Furthermore, it was used to design a novel metasurface featuring over 100% of relative bandwidth. Although the procedure has been used in the context of RCS design, it can be generalized to handle metasurface development for other application areas.

中文翻译：

---

基于机器学习的基于EM的框架，可有效，可靠地设计低散射超颖表面

由于其吸引人的特性，包括有效操纵电磁波（EM）的能力，超颖表面的受欢迎程度一直在持续增长。超颖表面包括布置为周期性晶格的晶胞的优化几何形状，以获得所需的EM响应。它们的新兴应用领域之一是隐身技术，尤其是降低雷达横截面（RCS）的实现。尽管有潜在的好处，但是缺乏系统的设计程序是阻碍超颖表面广泛应用的实际障碍。常规方法很大程度上受直觉启发，需要大量的设计人员进行交互，同时探索参数空间并追求最佳的晶胞几何形状。毫不奇怪，它们无法确定真正的最佳解决方案。在本文中，我们介绍了一种新颖的基于机器学习的框架，用于实现宽带RCS减少的超颖表面的自动化和高效计算设计。我们的方法是一个三阶段过程，涉及对晶胞进行全局替代辅助优化，然后对其局部进行优化。最后一个阶段是直接EM驱动的RCS减少带宽的最大化，这是通过适当制定涉及正则项的目标函数来实现的。所提出的框架的吸引人的特征在于，与仅优化单位单元的几何形状相反，它直接在整个超表面的水平上优化了RCS降低带宽。通过在前两个阶段进行高质量的初始设计，可以确保优化过程（尤其是最后阶段）的计算可行性。为了证实我们的程序的实用性，它已被应用于文献中报道的几种超表面设计，与原始设计相比，导致RCS减少带宽提高了15％–25％。此外，它还被用来设计一种新颖的超表面，其相对带宽超过100％。尽管该程序已在RCS设计的上下文中使用，但可以将其通用化以处理其他应用领域的超表面开发。